Pretreatment Nomogram for Predicting the Outcome
of Three-Dimensional Conformal Radiotherapy in
Prostate Cancer
By Michael W. Kattan, Michael J. Zelefsky, Patrick A. Kupelian, Peter T. Scardino, Zvi Fuks, and Steven A. Leibel
Purpose: Several studies have defined risk groups
for predicting the outcome after external-beam radio-
therapy of localized prostate cancer. However, most
models formed patient risk groups, and none of these
models considers radiation dose as a predictor vari-
able. The purpose of this study was to develop a nomo-
gram to improve the accuracy of predicting outcome
after three-dimensional conformal radiotherapy.
Materials and Methods: This study was a retrospec-
tive, nonrandomized analysis of patients treated at the
Memorial Sloan-Kettering Cancer Center between 1988
and 1998. Clinical parameters of the 1,042 patients
included stage, biopsy Gleason score, pretreatment
serum prostate-specific antigen (PSA) level, whether
neoadjuvant androgen deprivation therapy was ad-
ministered, and the radiation dose delivered. Biochem-
ical (PSA) treatment failure was scored when three
consecutive rises of serum PSA occurred. A nomogram,
which predicts the probability of remaining free from
NE OF THE FIRST nomograms for counseling pa-
O tients with localized prostate cancer, developed by
Partin et al,1 used clinical data of surgically treated patients
to predict the final pathologic stage after radical prostatec-
tomy. Pathologic stage had previously been shown to
correlate independently with each of three clinical variables,
clinical stage,2 Gleason score,3 and serum prostate-specific
antigen (PSA) level.4 Partin et al demonstrated increased
prediction accuracy when these variables were combined.
This nomogram was later revised and validated.5,6 How-
ever, because this nomogram did not predict other outcome
parameters, such as the risk of relapse, its clinical value has
been limited.7 Addressing this concern, Kattan et al8-10
subsequently developed nomograms that predict the proba-
bility of PSA relapse-free survival after radical prostatec-
tomy based on pre- and postsurgical variables. In the present
study, we extend this paradigm, providing a nomogram that
From the Departments of Urology, Epidemiology and Biostatistics,
and Radiation Oncology, Memorial Sloan-Kettering Cancer Center,
New York, NY, and the Cleveland Clinic, Cleveland, OH.
Submitted January 10, 2000; accepted May 12, 2000.
Address reprint requests to Michael W. Kattan, PhD, Memorial
Sloan-Kettering Cancer Center, 1275 York Ave C1068, New York, NY
10021; email kattanm@mskcc.org.
© 2000 by American Society of Clinical Oncology.
0732-183X/00/1819-3352
3352 Journal of Clinical Oncology, Vol 18, No 19 (October 1), 2000: pp 3352-3359
Downloaded from jco.ascopubs.org on September 27, 2010. For personal use only. No other uses without permission.
Copyright © 2000 American Society of Clinical Oncology. All rights reserved.
biochemical recurrence for 5 years, was validated in-
ternally on this data set using a bootstrapping method
and externally using a cohort of patients treated at the
Cleveland Clinic, Cleveland, OH.
Results: When predicting outcomes for patients in
the validation data set from the Cleveland Clinic, the
nomogram had a Somers’ D rank correlation between
predicted and observed failure times of 0.52. Predic-
tions from this nomogram were more accurate (P <
.0001) than the best of seven published risk stratifica-
tion systems, which achieved a Somers’ D coefficient of
0.47.
Conclusion: The development process illustrated
here produced a nomogram that seems to predict more
accurately than other available systems and may be
useful for treatment selection by both physicians and
patients.
J Clin Oncol 18:3352-3359. © 2000 by American
Society of Clinical Oncology.
predicts the probability of PSA relapse-free survival after
external-beam radiotherapy.
Several published models predict the outcome of radio-
therapy in localized prostate cancer.11-19 These models
combine prognostic variables to generate risk stratification
groups, based on initial clinical stage, Gleason score, and
pretreatment serum PSA levels. In general, these systems
consist of three to five risk groups and are designed to
predict the PSA relapse-free survival as a surrogate of tumor
control after radiotherapy. However, their predictive accu-
racy may be limited, because all of these systems combine
patients with similar, albeit not identical, variables to form
a risk group. The use of inhomogeneous groups potentially
results in suboptimal predictive accuracy. For example,
most models consider the range of PSA values of ⱕ 10
ng/mL as a uniform prognostic variable. However, a patient
with a PSA level of 4.1 ng/mL may not have the same
prognosis as the patient with a PSA level of 9.9 ng/mL, even
if other variable values are identical in a given case. Ideally,
the additional risk to a patient with a higher PSA value
should be considered when judging his risk of treatment
failure. A further limitation of previous models is that they
did not adjust for the radiation dose delivered. For example,
a recent outcome study of radiotherapy in prostate cancer
that pooled patients from several institutions14 did not
consider dose as a variable in the outcome analysis despite
the fact that the patients were treated with doses ranging
PRETREATMENT NOMOGRAM 3353

between 63 and 79 Gy. Recent studies demonstrated that Table 1. Descriptive Statistics for the Cohorts
treatment dose within this range significantly impacts the Nomogram Nomogram
Development Validation
outcome of radiotherapy in localized prostate cancer.15,20 (MSKCC) (Cleveland Clinic)
Our research question was whether a nomogram, which is No. of No. of
based on a continuous risk estimation model and includes Variable Patients % Patients %
the effect of dose as a variable, would improve the outcome Clinical (T) stage
prediction accuracy after radiotherapy in localized prostate T1c 276 26.5 302 33.1
cancer, using PSA relapse-free survival as an end point. To T2a 138 13.2 249 27.3
T2b 199 19.1 157 17.2
achieve this goal, several modeling techniques were tested
T2c 189 18.1 104 11.4
on a large cohort of patients with prostate cancer treated at T3ab 89 8.5 62 6.8
the Memorial Sloan-Kettering Cancer Center (MSKCC) T3c 151 14.5 38 4.2
with three-dimensional conformal radiotherapy (3D-CRT), Gleason sum
which is considered state-of-the-art in the radiation man- 2 4 0.4 3 0.3
3 27 2.6 6 0.7
agement of prostate cancer. The Cox proportional hazards
4 63 6.0 36 4.0
regression method21 with restricted cubic splines22 was 5 137 13.1 120 13.2
found to be the most accurately predicting technique and 6 327 31.4 367 40.2
thus was used to construct the nomogram. This nomogram 7 310 29.8 265 29.1
seemed to yield the most accurate prediction of the outcome 8 114 10.9 91 10.0
9 52 5.0 23 2.5
when tested on the MSKCC database and confirmed using a
10 8 0.8 1 0.1
database of patients with prostate cancer treated at the Neoadjuvant hormones
Cleveland Clinic. The proposed nomogram significantly No 655 62.9 705 77.3
improves the prediction accuracy when compared with Yes 387 37.1 207 22.7
several published risk stratification models. Dose, Gy
Minimum 64.8 64.8
1st quartile 70.2 68.4
MATERIALS AND METHODS Median 75.6 70.0
Mean 74.8 71.5
MSKCC Series 3rd quartile 75.6 74.0
Maximum 86.4 78.0
From December 1988 through November 1998, 1,080 patients with
Pretreatment PSA, ng/mL
prostate cancer were treated at MSKCC with 3D-CRT. Patients with Minimum 0.3 0.4
missing data (PSA, n ⫽ 2; Gleason score, n ⫽ 9; recurrence status, n ⫽ 1st quartile 6.7 6.5
27) were excluded, leaving 1,042 patients available for the present Median 11.0 9.9
analysis. The median age was 68 years (range, 46 to 86 years), and 94% Mean 17.9 16.9
of the patients were white. The clinical characteristics of this group of 3rd quartile 18.9 18.0
patients are listed in Table 1. Each patient was clinically staged by one Maximum 560.0 692.9
of three of the authors (M.J.Z., Z.F., S.A.L.) according to the 1992
American Joint Committee on Cancer staging classification system.23
Those with T3a and T3b disease were combined into a single group
because of the small frequencies of each. Serum PSA was measured by The parameters selected as nomogram variables included the clinical
the Tosoh radioimmunoassay, with a normal range of ⱕ 4.0 ng/mL, and stage, biopsy Gleason score, pretreatment serum PSA level, whether
modeled as its natural log. All patients had histologically confirmed neoadjuvant androgen deprivation therapy was administered, and the
adenocarcinoma of the prostate, classified according to the Gleason radiation dose delivered. Variables that are not available before
grading system3 by one of two pathologists at our institution. Radiation treatment, such as time to PSA nadir, or that do not have theoretical
was planned and delivered using the MSKCC system for 3D-CRT or justification for independent predictive ability, such as patient age,
the intensity-modulated version of 3D-CRT, as previously described.24 were not included in the nomogram. Treatment failure was defined
Table 1 shows that since 1991, 387 patients (37.1%) with large-volume according to a modification14 of the American Society of Therapeutic
prostate glands, in whom the treatment-planning dose-volume histo- Radiation Oncology consensus definition.25 It was based on the
grams indicated that greater than 30% of the rectal wall and/or greater development of three consecutive rises of serum PSA level. The date of
than 50% of the bladder wall might receive a dose of greater than 75 failure was defined as the midpoint between the last nonrising and the
Gy, were treated with neoadjuvant androgen deprivation for 3 months first rising PSA values. All patients were included with no minimum
before and during radiation.15 Hormone therapy was discontinued at requirements for follow-up or number of PSAs. None of the patients
the completion of radiotherapy. Radiation doses ranged between 64.8 received postirradiation hormonal or other anticancer therapy before
and 86.4 Gy, delivered in daily dose fractions of 1.8 Gy, using a documentation of PSA failure. The median follow-up time for the
previously reported dose-escalation scheme that achieved a dose of 81 patients who did not relapse was 29 months (range, 6 to 113 months).
Gy in October 1992.15 Follow-up evaluations were performed at 3- to The overall freedom from recurrence is illustrated in Fig 1. Removing
6-month intervals. from analysis the patients who were treated within the last 24 months14

Downloaded from jco.ascopubs.org on September 27, 2010. For personal use only. No other uses without permission.
Copyright © 2000 American Society of Clinical Oncology. All rights reserved.
3354
affects the probability of remaining free from recurrence by less than
1%, so these patients were retained in the analysis. There were no
recurrences beyond 63 months, and there were 649 patient-months of
follow-up beyond 60 months.
Validation Data Set: The Cleveland Clinic Series
The Cleveland Clinic data set included 1,030 patients treated
between July 1986 and November 1998. Of these, 78 patients who
received planned adjuvant hormonal therapy and another 40 in whom
one or more predictor variables and/or follow-up information was
missing were excluded from the current analysis. Of the remaining 912
patients, 529 received conventional external-beam radiotherapy and
383 received 3D-CRT. A Cox proportional hazards regression analysis
did not disclose any adjusted impact of the treatment technique
(conventional v 3D-CRT) on the therapeutic outcome in this series
(P ⫽ .14). Therefore, the entire data set of the 912 patients was used for
analysis in the present study. None of the patients received hormonal or
other anticancer therapy before the demonstration of recurrence. The
clinical characteristics of this group of patients are listed in Table 1, and
overall freedom from recurrence is illustrated in Fig 1. The median
follow-up time for the patients who did not recur was 25 months (range,
6 to 148 months).
Statistical Methods
In the initial phase of nomogram development, eight prediction
techniques were compared, including the Cox proportional hazards
regression method,21 Cox regression with restricted cubic splines to
allow nonlinear effects,22 an artificial neural network adapted for
censored data,26 a neural network for censored data free of the
proportional hazards assumption,27 recursive partitioning adapted to
censored data,26 a second recursive partitioning approach,28 a tree-
structured method for survival data,29 and multiple adaptive regression
Downloaded from jco.ascopubs.org on September 27, 2010. For personal use only. No other uses without permission.
Copyright © 2000 American Society of Clinical Oncology. All rights reserved.
KATTAN ET AL
Fig 1. Kaplan-Meier esti-
mates of disease-free probability
for the MSKCC and Cleveland
Clinic series of 1,042 and 912
patients, respectively. Numbers
above the months indicate pa-
tients at risk for recurrence
within each series. Abbreviation:
CC, Cleveland Clinic.
splines30 adapted to censored data.26 When the most accurate technique
was identified, it was represented graphically as a nomogram to
facilitate use for future prediction. This nomogram was evaluated using
bootstrapping31 to obtain relatively unbiased estimates of expected
future performance. This is a procedure whereby the data set is
repeatedly sampled and the nomogram regenerated and tested. The
nomogram was also validated externally using the Cleveland Clinic
series by comparing nomogram predictions with those produced by
seven risk stratification schemes from the literature. Although there are
several measures of predictive accuracy, we chose to use Somers’ D
rank correlation between predicted and observed failure times based on
the arguments made by Harrell et al.32 In this measure, a correlation
coefficient of 0 represents no discriminating ability and a value of 1
represents perfect discrimination. It can be converted to a concordance
index by dividing by 2 and adding to 0.5. Statistical significance was
assessed using the McNemar test for a difference in the frequency of
discordant pairs.33 All tests of statistical significance were two-sided.
All analyses were conducted with S-Plus software version 4.5 Profes-
sional Edition (Mathsoft Data Analysis Products Division, Seattle,
WA) with the Design and Hmisc libraries.22
RESULTS
When the prediction techniques were compared, the Cox
proportional hazards model with restricted cubic splines
produced the most accurate predictions. This Cox prediction
model is illustrated graphically by the nomogram in Fig 2.
Using bootstrapping to obtain unbiased estimates of ex-
pected future performance, this nomogram predicted the
5-year PSA relapse-free survival with a Somers’ D corre-
lation coefficient of 0.46, and the performance across the
spectrum of relapse probabilities was accurate to within
PRETREATMENT NOMOGRAM
Fig 2. Radiotherapy nomogram
based on 1,042 patients treated at
MSKCC for predicting PSA recur-
rence after radiation therapy.
10% of the observed freedom from PSA relapse (Fig 3). The
predictive accuracy of this nomogram was compared with
the accuracy of the seven risk stratification schemes using
the Cleveland Clinic data (Fig 4). For this comparison, each
patient in the Cleveland Clinic series had his outcome
independently predicted by the nomogram and each of the
risk stratification schemes, and all of these predictions were
compared with observed outcomes. The McNemar test
indicated that the Somers’ D correlation coefficient of the
nomogram (0.52) was higher (P ⫽ .0001) than that of the
best risk stratification scheme tested (0.47), suggesting that
the nomogram was more accurate than the risk stratification
schemes in predicting the outcome.
DISCUSSION
The present study demonstrates that the nomogram de-
veloped here represents a valid and accurate method for
predicting the outcome of external-beam radiotherapy in
localized prostate cancer. Its advantage over the risk strat-
ification prediction systems is likely associated with the use
of continuous risk scales for relevant technique parameters
Downloaded from jco.ascopubs.org on September 27, 2010. For personal use only. No other uses without permission.
Copyright © 2000 American Society of Clinical Oncology. All rights reserved.
3355
as opposed to condensing sections of the risk spectra into
heterogeneous risk groups. The inclusion of dose as a
variable affecting outcome may have also contributed to the
improved predictive accuracy. Definitive interpretations
regarding the degree of the improvement in prediction are,
however, difficult because the data are of the time-until-
event type. Thus, although the nomogram was shown to
predict better than the best of the risk stratification models
(P ⬍ .05), the improvement is difficult to appreciate by
interpreting the error measure used in this study, the
Somers’ D correlation coefficient, which evaluates the
strength of rank association between a prognostic model and
the actual PSA, relapse-free survival. To address this
concern, it is helpful to illustrate the use of the nomogram
and compare its prediction with that of risk stratification by
considering a specific patient from the Cleveland Clinic
validation series. This patient had clinical stage T2c disease,
a pretreatment PSA level of 6 ng/mL, a biopsy Gleason
score of 9, and a planned dose of 66.6 Gy without use of
neoadjuvant hormones. By a recently published risk strati-
fication scheme,14 this patient would be classified into the
3356
Fig 3. Nomogram Calibration. Horizontal axis is nomogram prediction.
Vertical axis is actual PSA recurrence-free survival at 60 months. Dashed line
represents an ideal nomogram. Solid line is current nomogram performance
with 95% confidence intervals. E represents data set subcohorts. ⴛ repre-
sents the bootstrap-corrected estimate of nomogram.
most favorable group (group I), which has an associated
freedom from recurrence prediction of 81% at 5 years.
According to our nomogram, which was shown to predict
more accurately on the Cleveland Clinic series, the same
patient’s predicted probability of 5-year freedom from
recurrence is 24%. With our previously reported preopera-
tive surgery nomogram,8 the same patient’s predicted prob-
ability of 5-year freedom from recurrence is 68%. Thus, risk
stratification may have influenced this man to choose
radiation therapy over surgery, whereas with the nomogram
approach, this particular patient’s better chance at freedom
from progression would seem to have been with surgery
Fig 4. Comparison of nomogram with risk stratification models (desig-
nated by bibliographical reference number) using Cleveland Clinic data.
Horizontal axis is Somers’ D. Notice that the nomogram is more accurate
than the risk stratification models (P < .0001).
Downloaded from jco.ascopubs.org on September 27, 2010. For personal use only. No other uses without permission.
Copyright © 2000 American Society of Clinical Oncology. All rights reserved.
KATTAN ET AL
rather than with low-dose radiation. As a second example,
consider the Cleveland Clinic patient who had T1c disease
with a PSA level of 21.4 and a biopsy Gleason sum of 7 who
was treated with 78 Gy and neoadjuvant hormones. Risk
stratification14 would have put him in the least favorable
risk group, with a 29% 5-year freedom from recurrence
prediction. The surgery nomogram8 predicts a 55% freedom
from recurrence, whereas the radiation therapy nomogram
predicts 82% freedom from recurrence, both at 5 years.
Thus, risk stratification would have potentially directed this
patient away from radiation therapy, which seems to be his
better choice for preventing disease progression. Figure 5
compares the nomogram and risk stratification14 predictions
for all patients in the Cleveland Clinic series. It is clear from
this Figure that the risk stratification approach produces
heterogeneous strata with respect to the nomogram predic-
tions and, assuming the nomogram is indeed more accurate,
could be misleading for the individual patient who is trying
to decide on a treatment strategy. The latter is illustrated by
the number of patients with low freedom from recurrence
predictions by the nomogram despite favorable risk strati-
fication and the number of patients with high freedom from
recurrence predictions by the nomogram despite unfavor-
able risk stratification. Thus, risk stratification does not
seem to provide these patients with the best estimates of
treatment efficacy currently available and may lead to them
making poor treatment choices as they compare treatments.
The Somers’ D method uses a range of coefficient values
from ⫺1 to ⫹1 to test the predictive accuracy associated
with a model. On this scale, 0 represents no association at
all and 1 (or ⫺1) represents perfect positive (or negative)
association. The nomogram developed here performs near
the center of this scale, discriminating significantly better
than chance (P ⬍ .0001). Although it provides improved
accuracy over the risk stratification models we tested, all of
which had smaller D scores, it leaves a substantial degree of
uncertainty in the prediction. It seems that an improved
nomogram will have to be based on larger patient series,
longer follow-up periods, and additional biologic and clin-
ical markers with significant predictive potential. Although
the follow-up in the validation data set was not as mature as
that found in the nomogram derivation data set, Somers’ D
is able to use all follow-up information, regardless of
duration. Shorter follow-up in a series does not bias one
prediction technique in favor of another; it simply reduces
the statistical power to declare one technique superior to
another. Fortunately, sample size and follow-up were ade-
quate in the validation data set to provide sufficient power to
declare one technique superior.
To identify the best algorithm for nomogram design, we
compared several statistical and machine learning models of
PRETREATMENT NOMOGRAM
Fig 5. Nomogram predic-
tions within each level of risk
stratification, I (most favorable)
through IV (least favorable).14
Within each risk group quadrant
is a histogram of the nomo-
gram-predicted probabilities (x-
axis) of freedom from recur-
rence.
outcome prediction. This trial-and-error approach was cho-
sen because it was impossible to know in advance which
method would provide the best prediction on a given set of
data, because each technique has theoretical strengths and
weaknesses.34 Of the techniques considered, the traditional
Cox model is the most commonly used method for multi-
variable analysis of survival data. It assumes that continuous
variables have linear effects (ie, linearly related to the
log-hazard function), unless special modifications are made,
such as restricted cubic splines. In fact, in the present study
we used the restricted cubic splines approach for the PSA,
Gleason, and dose variables, which improved the predictive
accuracy (data not shown). This study was not an exhaustive
comparison of the modeling techniques tested. Each tech-
nique has numerous options that may be varied, and most
options were left at the software defaults. Thus, our results
do not represent definitive comparisons of the techniques in
a general sense, nor can it necessarily be concluded that the
Cox model with restricted cubic splines is the best tool for
the analysis of our data. Because a definitive comparison of
techniques was not the focus of this study, we did not
conduct statistical significance testing between techniques
but merely selected the technique with the superior apparent
accuracy. A thorough comparison of all the options for each
technique would require an enormous study. Simply put, we
Downloaded from jco.ascopubs.org on September 27, 2010. For personal use only. No other uses without permission.
Copyright © 2000 American Society of Clinical Oncology. All rights reserved.
3357
carried out limited comparisons of each technique and
found the Cox model with restricted cubic splines to be the
most accurate. Others may find, by varying the options, that
a different approach may further improve the predictive
accuracy of the nomogram. Nonetheless, the Cox model
with restricted cubic splines produced a nomogram that
predicts more accurately than the risk stratification systems
previously used for outcome prediction. We plan to facili-
tate the computations of the nomogram by adding it to our
Palm Pilot (Palm, Inc, Santa Clara, CA) nomogram software
application, which we distribute free of charge.9
In addition to serving as a prognostic tool, the nomogram
in Fig 2 is also useful in interpreting the underlying Cox
model. PSA seems to have a major impact across its
spectrum. The clinical stage point assignment is intuitive
except for, possibly, stage T2c, which is associated with a
worse predicted prognosis than is T3ab. The difference
likely results from estimation variability for the coefficient
or variability in the digital rectal examination. However, the
difference between these groups is not statistically signifi-
cant at the 5% level. We have avoided combining groups
after modeling to improve the appearance of the nomogram
because such action can have a deleterious effect on
performance.32 This is also the reason for not deleting the
nomogram axis for “Hormones,” which has a statistically
3358
insignificant impact on the outcome. When interpreting the
nomogram, it is essential to consider possible changes in
other variables (eg, PSA level and Gleason score) when
comparing points across levels of a single variable (eg,
dose). It is difficult to draw meaningful conclusions about
the effect of a single variable in isolation when other
variables of the nomogram may correlate with it, because
moving a patient on one axis would likely affect the position
of variables on other axes. Thus, whereas the nomogram
indicates that dose may be a helpful predictor of patient
outcome, this analysis is not a formal test of dose-response.
It is also important to consider our selection criteria for
neoadjuvant hormones when judging its effect on predicted
probability.
There are several potential applications for nomograms
such as that developed in the present study.8,9 First, it may
assist both physicians and patients in treatment selection. A
patient weighing the risks and benefits of various treatments
would value the most accurate predictions of treatment
outcome currently available when he considers radiation
therapy. It should, however, be emphasized that the devel-
opment of the present nomogram was based on a cohort of
patients treated with external-beam radiotherapy. Both phy-
sician and patient biases unquestionably affected the selec-
tion of treatment in this group of patients. Hence, the
nomogram parameters may not be applicable to the popu-
lation of patients with localized prostate cancer at large. It
should, nonetheless, be noted that approximately 75% of the
patients in the MSKCC series were stage T1 or T2 patients
and that a majority of these patients were offered a surgical
option. Whether simultaneous use of surgical and radiation
nomograms would facilitate treatment selection in an indi-
vidual candidate remains to be tested.
The design of clinical trials may represent another appli-
cation of nomograms. By quantifying the probability of
treatment failure, prognostic nomograms also identify pa-
tients most likely to benefit from neoadjuvant or adjuvant
treatments. Selection of candidates would be facilitated by
specifying a risk cutoff (eg, at least 70% chance of treatment
failure) rather than individual variable levels, which may
not as easily produce homogenous groups of patients. A
third use, also related to clinical trials, may be verification
REFERENCES
1. Partin AW, Yoo J, Carter HB, et al: The use of prostate specific
antigen, clinical stage and Gleason score to predict pathological stage
in men with localized prostate cancer. J Urol 150:110-114, 1993
2. Jewett HJ, Bridge RW, Gray GF Jr, et al: The palpable nodule of
prostatic cancer: Results 15 years after radical excision. JAMA
203:403-406, 1968
Downloaded from jco.ascopubs.org on September 27, 2010. For personal use only. No other uses without permission.
Copyright © 2000 American Society of Clinical Oncology. All rights reserved.
KATTAN ET AL
of adequate randomization. Traditionally, such verifications
have been carried out by comparing the distribution of
individual variables in each arm and recording differences.
However, joint distribution differences may remain unde-
tected by this procedure. Assessment of a predicted risk
imbalance using all of the common prognostic factors
would yield more relevant comparisons.
There are several limitations to the present nomogram.
The major limitation is the high level of uncertainty in
prediction that still exists. In addition to the fact that the
bootstrap-corrected Somers’ D correlation coefficient for
the nomogram was 0.46, the confidence intervals at the
various predicted probabilities of recurrence (Fig 3) were at
some levels as high as ⫾ 10%. Furthermore, as we state on
the nomogram itself, disease may still recur in patients after
the prediction time point of 5 years, although this has been
shown to be rare.35 However, although the current nomo-
gram may not represent the most optimal tool for prediction,
it nonetheless provides an example of the appropriate
methodology for development and evaluation of prognostic
nomograms in general. Furthermore, the Somers’ D coeffi-
cient of 0.46 is near that of other prognostic models of
survival data.36 Another limitation is that the nomogram is
based on a single institution’s database and that the patients
were treated with the sophisticated and technically ad-
vanced 3D-CRT technique. Despite the fact that it was
validated by the Cleveland Clinic series of mixed modalities
with dose escalation, its applicability to other centers and
external-beam radiotherapy of localized prostate cancer in
general require further testing. We encourage others to
evaluate the accuracy of this nomogram and development
approach on their series. The fact that patient evaluation
with this nomogram encompasses commonly used pretreat-
ment parameters, the planned dosage, and the decision
regarding the frequently applied neoadjuvant androgen
deprivation suggests that its usefulness in outcome predic-
tion of radiotherapy in stages T1c to T3c NX MO prostate
cancer can be rapidly assessed. It may provide a highly
beneficial tool for physicians and patients in making deci-
sions about treatment and in identifying patients at high risk
of failure after radiotherapy who may benefit from adjuvant
treatment protocols.
3. Gleason DF: Histologic grading and clinical staging of prostatic
carcinoma, in Tannenbaum M (ed): Urologic Pathology: The Prostate.
Philadelphia, PA, Lea and Febiger, 1977
4. Stamey TA, Yang N, Hay AR, et al: Prostate-specific antigen as
a serum marker for adenocarcinoma of the prostate. N Engl J Med
317:909-916, 1987
PRETREATMENT NOMOGRAM
5. Partin AW, Kattan MW, Subong ENP, et al: Combination of
PSA, clinical stage and Gleason score to predict pathological stage in
men with localized prostate cancer: A multi-institutional update. JAMA
277:1445-1451, 1997
6. Blute ML, Bergstralh EJ, Partin AW, et al: Validation of the
Partin tables for predicting pathologic stage of clinically confined
prostate cancer. 94th Annual Meeting of the American Urological
Association, Dallas, TX. J Urol 161:238, 1999 (abstr)
7. Kattan MW, Stapleton AMF, Wheeler TM, et al: Evaluation of a
nomogram for predicting pathological stage of men with clinically
localized prostate cancer. Cancer 79:528-537, 1997
8. Kattan MW, Eastham JA, Stapleton AMF, et al: A preoperative
nomogram for disease recurrence following radical prostatectomy for
prostate cancer. J Natl Cancer Inst 90:766-771, 1998
9. Kattan MW, Wheeler TM, Scardino PT: A postoperative nomo-
gram for disease recurrence following radical prostatectomy for pros-
tate cancer. J Clin Oncol 17:1499-1507, 1999
10. Kattan MW, Witte M, Morton RA, et al: Validation of a nomogram
for predicting disease recurrence following radical prostatectomy for
clinically localized prostate cancer. 94th Annual Meeting of the American
Urological Association, Dallas, TX. J Urol 161:329, 1999 (abstr)
11. D’Amico AV, Whittington RS, Malkowicz SB, et al: Pretreat-
ment nomogram for prostate-specific antigen recurrence after radical
prostatectomy or external-beam radiation therapy for clinically local-
ized prostate cancer. J Clin Oncol 17:168-172, 1999
12. Zagars GK, Geara FB, Pollock A, et al: The T classification of
clinically localized prostate cancer: An appraisal based on disease
outcome after radiation therapy. Cancer 73:1904-1912, 1994
13. D’Amico AV, Whittington R, Malkowicz SB, et al: Biochemical
outcome after radical prostatectomy, external beam radiation therapy,
or interstitial radiation therapy for clinically localized prostate cancer.
JAMA 280:969-974, 1998
14. Shipley WU, Thames HD, Sandler HM, et al: Radiation therapy
for clinically localized prostate cancer. JAMA 281:1598-1604, 1999
15. Zelefsky MD, Leibel SA, Gaudin PB, et al: Dose escalation with
three-dimensional conformal radiation therapy affects the outcome in
prostate cancer. Int J Radiat Oncol Biol Phys 41:491-500, 1998
16. Mclean M, Panzarella T, Warde P, et al: A predictive model for
prostate cancer. Clin Invest Med 20:S90, 1997 (abstr)
17. Duchesne GM, Bloomfield D, Wall P: Identification of interme-
diate-risk prostate cancer patients treated with radical radiotherapy
suitable for neoadjuvant hormone studies. Radiother Oncol 38:7-12,
1996
18. Zagars GK, Pollack A, von Eschenbach AC: Prognostic factors
for clinically localized prostate carcinoma. Cancer 79:1370-1380, 1997
19. Pisansky TM, Kahn MJ, Bostwick DG: An enhanced prognostic
system for clinically localized carcinoma of the prostate. Cancer
79:2154-2161, 1997
Downloaded from jco.ascopubs.org on September 27, 2010. For personal use only. No other uses without permission.
Copyright © 2000 American Society of Clinical Oncology. All rights reserved.
3359
20. Hanks GE, Hanlon AL, Schultheiss TE, et al: Dose escalation with
3D conformal treatment: Five year outcomes, treatment optimization, and
future directions. Int J Radiat Oncol Biol Phys 41:501-510, 1998
21. Cox DR: Regression models and life tables. J Royal Stat Soc
34:187-220, 1972
22. Harrell FE: Design. S-Plus function for biostatistical/epidemio-
logic modeling, testing, estimation, validation, graphics, prediction,
and typesetting by storing enhanced model design attributes in the fit.
Http://lib.stat.cmu.edu/S/Harrell/Design.README
23. AJCC Cancer Staging Manual (ed 4). American Joint Commit-
tee on Cancer. Philadelphia, PA, Lippincott-Raven, 1992
24. Leibel SA, Zelefsky MJ, Kutcher GJ, et al: The biological basis
and clinical application of three-dimensional conformal external beam
radiation therapy in carcinoma of the prostate. Semin Oncol 21:580-
597, 1994
25. American Society for Therapeutic Radiology and Oncology
Consensus Panel: Consensus statement: Guidelines for PSA follow-
ing radiation therapy. Int J Radiat Oncol Biol Phys 37:1035-1041,
1997
26. Kattan MW, Hess KR, Beck JR: Experiments to determine
whether recursive partitioning (CART) or an artificial neural network
overcomes theoretical limitations of Cox proportional hazards regres-
sion. Comput Biomed Res 31:363-373, 1998
27. Ripley RM: survnnet: Neural network model for survival mod-
els. 1998. Http://www.stats.ox.ac.uk/pub/swin/survnet.zip
28. Atkinson E, Therneau T: rpart: Recursive partitioning and
regression trees. Http://lib.stat.cmu.edu/S/rpart
29. Segal MR: Regression trees for censored data. Biometrics
44:35-47, 1998
30. Kooperberg C, Bose S, Stone CJ: Polychotomous regression.
J Am Stat Assoc 92:117-127, 1997
31. Efron B, Tibshirani RJ: An Introduction to the Bootstrap. New
York, NY, Chapman and Hall, 1993
32. Harrell FE, Lee KL, Mark DB: Multivariable prognostic models:
Issues in developing models, evaluating assumptions and adequacy,
and measuring and reducing errors. Stat Med 15:361-387, 1996
33. Conover WJ: Practical Nonparametric Statistics (ed 2). New
York, NY, John Wiley and Sons, 1980
34. Kattan MW, Cooper RB: The predictive accuracy of computer-
based classification decision techniques: A review and research direc-
tions. Omega 26:467-482, 1998
35. Kestin LL, Vicini FA, Ziaja EL, et al: Defining biochemical cure
for prostate carcinoma patients treated with external beam radiation
therapy. Cancer 86:1557-1566, 1999
36. Knaus WA, Harrell FE, Lynn J, et al: The SUPPORT prognostic
model: Objective estimates of survival for seriously ill hospitalized
adults. Ann Intern Med 122:191-203, 1995